The number of partners and sub-partners reporting narrative data

As we have seen previously, Not all partners (and sub-partners) enter narrative records. For instance, the partner PMA has lots of sub-partners reporting for the different data types but there are no narratives there.

In terms of narrative data,

TODO The partners X and Y is like that. The rest of the main partners, namely … do not have any subpartners reporting by them.

• The partners do not have any sub-partners reporting via them: CARE, Dialogo Diverso, JRS Ecuador, Mision Scalabriniana, NRC, OIM, OPS/OMS, Plan Internacional, PNUD, UNESCO, UNFPA

The number cantons and provinces recording narrative data

Treemap plot showing canton and province reporting frequencies.

Word count

One issue with the nature of the questions is that they are only unique in a form. These questions can be distributed across multiple forms. The questions sharing the same name will have different meanings. For instance, the question “Cualitativo” from the form “Salud” should imply different thing than the question “Cualitativo” from the form “Protección_VBG”.

In order to solve this kind of problem:

• We can combine question with the form and also its folder label. There we can achieve a unique name for each question.

• Another thing to resolve this would be doing analysis to move the analysis up to form level. In this file, we did both, therefore the analysis shown as below:

Count of responses per topic/question:

It’s also a good practice to see the number of questions. For example, one question has two responses, therefore they’re short. Therefore, jittered points are added to give a glance about the number of observations in the same plot.

In the plot above, the box plot of form topics and response word counts based on the raw data, the outliers are shown in orange color. Outliers are the points placed outside the whiskers, which is the long line, of the boxplot.

The response word count distribution per form topic categorized by partner name:

The response word count distribution per form topic categorized by canton name:

A caveat: Reducing multiple values down to a single value should be avoided in the early stages of the analysis because reducing hides a lot e.g. a bar chart showing average the word count per partner. Some partners may write longer than others, because:

1. They actually write longer than other partners.

2. The questions they answered require short answers.

The Description field

Some questions have the description field giving extra details about the questions.

Do some questions with the extra description field have better response quality than the questions which do not have it?

Looking at the table containing form name, question, description and so on:

We see in the plot below that the response word counts per form and colored if a response has a description field or not. Having a description field or not is calculated as that a description field has a minimum one word.

The responses with the longest word counts are the ones with description. Nevertheless, it is not so easy to see a clear trend that there’s a correlation between response word count and description fields. Interestingly, the form topic Protección_VBG has no description fields at all in its form topics.

Analysis of Variance

TODO ANOVA

Trends

Correlation

TODO

The regression line

We can look at multiple continuous variables in our data.

• word count of response field: the dependent variable.

• word count of question field: an independent variable.

• word count of description field: an independent variable.

Scatter plots help understand the characteristics of those variables. However, we miss a general understanding that is the trend line.

The gray area around the lines shows the confidence band at the 0.95 level. Although there’s a straight slope in the linear regression line, we cannot say that the trend line is robust because the confidence band representing the uncertainty in the estimate is wide.

Logistic regression

TODO

Textual data preparation

Describe how to prepare textual data and what common steps are usually performed.

They are usually four steps involved in this process:

1. Tokenization

Tokenization means to split a text into tokens considered meaningful units of text. A token can either be a word (and often it is) or a group of words (such as bigram), or even a sentence that depends on the level of analysis.

Perform stemming, which you bring nouns/verbs back to infinitive forms, and tokenization, which is separating words in meaningful pieces.

2. Strip punctuation

Punctuation is often not required in text analysis (unless a researcher wants to tokenize the text based on a specific classifier such as sentence tokens); therefore, they create noise.

3. Convert text into lowercase

When the text turned into lowercase, for instance, the words respuesta and Respuesta will no longer be taken as different words.

4. Exclude stopwords & numbers

Stop words usually mean the most common words in a language that will bring no significant results in analysis. They are overly distributed in the text and they will not give so meaningful results itself. Stop-words are including articles (el/la), conjunctions (y), pronouns (yo/tú/etc.) and so on.

In text mining, this process is usually done after the text converted into lowercase so one does not have to provide stop words including both lower and sentence case versions.

We import a list of Spanish stopwords data (source here) and perform a filtering join returning tokens from textual data by excluding the words listed in the stopwords. that only returns the tokens not listed in the stopwords.

It’s also possible to add more custom words such as ACNUR or HIAS if such organization names are not desired in the results.

The original tokens had 37845 rows but after stop words, it decreased to 19221 and that the change in between is 51%.

The most common words

The most common words in all responses:

Sentiment analysis

Sentiment analysis (also called as opinion mining) is a technique to understand the emotional meanings of text given by a dictionary describing the positive/negative words that already done by humans.

The responses seem to be written with a formal tone of voice; therefore, the responses may not show any sentiment at all.

First, we find a sentiment lexicon for the Spanish language (source here).

A wordcloud showing positive and negative words: